Research On Technologies Of Image Enhancement And Edge Detection Based On Deep Convolutional Networks

According to MEMS Consulting,the overall market size of computer vision has reached 11.94 billion dollars at 2018,and it is expected to reach 17.17 billion dollars by 2023.Computer vision has become the most practical and promising technology in the field of artificial intelligence.At present,In various computer vision tasks,images are the basis of research.The quality of images determines the effect of image processing.However,due to the limitations of the hardware and the impact of the environment and the proficiency of oprators,the quality of images often does not meet the actual requirements,especially the limitation of resolution and the blurring of images during the shooting process.These factors may caused severe interference to subsequent computer vision tasks,limit the practical application of computer vision,have become a bottleneck restricting the further development of deep learning technology.Image enhancement has become a research hotspot in the field of artificial intelligenceThe imaging process will be disturbed by various factors in reality,which will affect imaging quality.Computer vision technologies such as image enhancement are of vital importance to imaging problems in reality.For instance,in the field of chemical synthesis,scanning electron microscopy is a common tool to explore the morphological changes during the growth of nanocrystals,to prepare new materials.However,the imaging quality of SEM is affected by the optical system and the operator.Obtaining high-resolution SEM images with clear edges and rich details is an urgent issue.In addition,there are lots of information unrelated to the crystal structure in SEM images.Such as substrate background and other products,bad for the direct observation of the crystal.Regarding the issues above,in this paper we explores image enhancement methods including image super-resolution(SR)and image deblurring,and image edge detection methods,based on deep convolutional networks such as residual networks and generative adversarial networks,to explore the improvement of image enhancement on edge detection.And taking SEM image enhancement and edge detection as examples,we explored the practical significance and value of the method in real scenarios.The main innovations are as follows1.An image SR reconstruction method based on deep residual network and dense connection is proposed.We use dense connections to improve the residual structure and enhance the network's feature propagation and reuse.During training,we use the Nadam optimizer to improve the training effect and convergence.Good results have been obtained on the DVI2K data set.EDSR is one of the deep residual networks with the best SR effect at present.Our method improves 0.0156 over EDSR on SSIM,and has better visual effects.The SR reconstruction of SEM images also has better results.Compared with EDSR,PSNR and SSIM are improved by 0.25dB and 0.03.2.An image deblurring method based on denormalized GAN network is proposed.DeblurGAN is a groundbreaking deblurring method based on GAN.We optimize it in:all normalization layers in generator are removed to eliminate the contrast stretching problem caused;we use more residual blocks to increase the depth of the generator and use larger receptive fields to improve the effect of generation.The resluts on the GOPRO dataset proves that our method improves 0.35dB and 0.0023 over DeblurGAN on PSNR and SSIM,and reduces the processing time of a single image by 11.5%.On the deblurring of SEM images,our method has an improvemnet of 1.4%and 1.6%than DeblurGAN on Brenner and Tenengrad,the results have richer details and higher clarity.3.An edge detection method based on residual network and multi-scale features is proposed.The residual network is introduced to improve the convergence of the network and the stability of training.Subpixel convolution is used as an upsampling method for scaling the feature map Compared with Canny and current mainstream methods such as HED,RCF,our method has better results.The F-measure on BSDS500 dataset has increased by 10.81%,5.34%and 1.07%;And it also has better continuity and accuracy in contour detection of nanocrystal SEM images,and can eliminate background noise.We performed edge detection on the SEM image after super-resolution and deblurring reconstruction.Maintaining the original edge continuity and accuracy,it further eliminated the background noise and detected finer edges,which proved the image enhancement has a certain improvement effect on edge detection.